Abstract: Soil degradation is often reflects grassland degradation. Monitoring soil organic carbon (SOC) content over large areas using remote sensing technology can help assess soil condition allowing better understanding of grassland degradation. Alpine grassland in the Gannan Prefecture was selected for this research. NIR-Visible spectral reflectance of grassland soil samples was measured using ASD (analytical spectral devices) spectroradiometer under laboratory conditions. Correlation analyses between eight transformations of soil spectral reflectance and SOC content were undertaken and sensitive wavebands selected. Three multivariate regression techniques (stepwise multiple linear regression, SMLR, principal components regression, PCR, partial least squares regression, PLSR) were compared with the aim of identifying the best inversion model to estimate alpine grassland SOC. The determination coefficient of validation dataset (R_v²), the root mean square error (RMSE), and the residual prediction deviation (RPD) were used to evaluate the models. The result indicated that differential transformation could improve the correlation between spectral characteristics and SOC content. The first derivative of reflectance had the best correlation with SOC content during transformation, the maximum correlation coefficient value was 0.865; Three multivariate regression models based on the first derivative of reflectance had excellent SOC prediction capability and recommended as a good spectral transformation for reliably estimating SOC. Comparing the multivariate regression techniques based on all transformations, PLSR performed best (high R_v² and RPD, low RMSE), which suggests that PLSR is the most suitable method for estimating SOC content in this study. The optimal SOC estimation model of Gannan alpine grassland was the combination of PLSR and the first derivative of log reflectance spectra [(lgR)'], providing a relatively high coefficient of determination for the validation set (R_v²=0.878), the highest residual prediction deviation (RPD=2.946) and the lowest root mean square error (RMSE=7.520). The RPD of the optimal model was higher than 2.5, which suggested that the model was robust and stable enough to be applied for estimation of SOC in other areas.